Anthropic released Claude Opus 5.5 on 22 September 2026 at $4 per million input tokens and $20 per million output, down from Opus 5's $5 and $25. A new flagship that costs less than the one it replaces is rare enough to be the headline on its own.
What changed
Opus 5.5 is the first model in Anthropic's 5.5 family. The list price is a 20% cut against Opus 5. Cache reads dropped further, from $0.50 to $0.20 per million tokens, a 60% reduction, which matters more than it sounds for any workload that re-sends the same context repeatedly.
Anthropic reports Opus 5.5 topping its benchmark table on SWE-bench Pro at 89.9% and Terminal-Bench 4.0 at 66.4%. The company also cites a practical test: porting a legacy HAProxy codebase from C to Rust in 9.5 hours, against 12 hours for Claude Fable 5.1, a 51% improvement in throughput terms.
Take vendor benchmarks as vendor benchmarks. They are run by the people selling the model, on tasks they chose. The direction is almost certainly right; the precise margins are worth less than your own testing.
The Fable question
Anthropic now sells two models above Sonnet: Opus 5.5 at $4/$20 and Fable 5.1 at $10/$50. Before this release the ordering was simple, with Fable at the top and Opus below it. Opus 5.5 muddies that deliberately.
On cost per message the gap is wider than the per-token gap suggests, because Fable tends to produce longer answers. A 3,000-token prompt with a 1,500-token response costs about $0.042 on Opus 5.5 and about $0.105 on Fable 5.1. If Fable writes twice as much, which it often does, the real ratio is over four times.
Our read: start on Opus 5.5 and escalate to Fable only when you have a specific reason. Long-horizon agentic work and the hardest coding problems are still where Fable earns its price. General frontier work is not, any more.
What stayed the same
Opus 5.5 keeps the 1,000,000-token context window and the 128,000-token maximum output. Input modalities are text, images and files. If you were already building against Opus 5, nothing about the envelope forces a rewrite.
It is available through Anthropic's API and across AWS, Google Cloud and Microsoft's ecosystem, and it is on Deepest as of the day it launched.
How to actually evaluate it
The useful test is not SWE-bench. It is whether Opus 5.5 handles the specific thing you keep needing a frontier model for, and whether it does so well enough that the cheaper tiers stop being tempting.
The fastest way to find out is to put the same prompt in front of Opus 5.5, Fable 5.1 and a cheaper model at once and look at the three answers together. Half the time the cheap model is fine and you have been overpaying. The other half you see exactly what the extra money buys. Either result is worth knowing, and neither is visible if you only ever ask one model.
Frequently Asked Questions
Should I switch from Opus 5 to Opus 5.5?
For most work, yes: it is the newer model and it costs less per token. The only reason to hold is if you have evaluation results specific to Opus 5 that you would need to redo, which is a real cost for production systems and not one to dismiss.
Is Opus 5.5 better than GPT-6 Sol or Grok 4.7?
Nobody can answer that for your work, including us. All three landed within 48 hours of each other at prices close enough that cost is not the deciding factor. On Deepest you can ask all three the same question and compare the answers directly, which is a more honest answer than a benchmark table.
Does the cache read price cut matter?
If you re-send the same long context repeatedly, such as a large codebase or document set across many turns, dropping from $0.50 to $0.20 per million cached tokens is a larger saving than the headline price cut. If your prompts are short and varied, it changes almost nothing.
